Built a mini-CNN framework with numeric gradient checks; fine-tuned ResNet18 on EMNIST, reaching 96.5% test accuracy.
2025
Tech stack
Problem
Implemented a convolutional neural network from scratch in PyTorch with numeric gradient verification, then fine-tuned ResNet18 on EMNIST using a staged linear-probe-to-full-finetune curriculum.
Approach
Ran on SLURM HPC with automated ablation scripts. Achieved 96.5% test accuracy..
Result
See the code for full results.
Full writeup
Implemented a convolutional neural network from scratch in PyTorch with numeric gradient verification, then fine-tuned ResNet18 on EMNIST using a staged linear-probe-to-full-finetune curriculum. Ran on SLURM HPC with automated ablation scripts. Achieved 96.5% test accuracy.